RoboCOIN

A cross-embodiment bimanual manipulation dataset led by BAAI, open-sourced across 15 kinds of robots.

RoboCOIN is a bimanual-manipulation dataset led by the Beijing Academy of Artificial Intelligence (BAAI), open-sourced in November 2025 together with more than twenty universities and robotics companies. It contains over 180,000 demonstrations from 15 robot platforms, including bimanual robots, half-body humanoids, and full-body humanoids (models from manufacturers such as AgileX, Galaxea, AgiBot, Galbot, Leju Robotics, Unitree, and others), covering 16 environments, 421 bimanual tasks, and 432 objects, with tasks organized into 39 categories of bimanual coordination behavior. Annotation is done at three levels: a task concept for the whole trajectory, segmented subtasks, and frame-by-frame kinematic information. A companion processing pipeline, CoRobot, built on LeRobot, handles data quality control, automatic annotation, and unified management of cross-embodiment data.

ExampleA single bimanual demonstration carries three layers of annotation: the whole trajectory states what task is being done, segments mark the start and end of each subtask, and each frame carries kinematic data; the data can be loaded directly for training using tools from the LeRobot ecosystem.

Also called

RoboCoin, RoboCOIN: An Open-Sourced Bimanual Robotic Data Collection for Integrated Manipulation

Related

Beijing Academy of Artificial Intelligence · Bimanual Manipulation · Cross-Embodiment Data · LeRobotDataset · Subtask Segmentation · AgiBot World

Sources

RoboCOIN (arXiv:2511.17441)
RoboCOIN 项目主页 (Chinese)

As of

2026-04